The same shape in a rising and in a falling market is a different situation, even when normalisation lays the curves on top of each other. A bounce inside a multi-year uptrend and an identically shaped bounce inside a crash have little in common beyond the shape.
So when historical analogues are searched for, shape is not compared on its own but together with the regime it happened in.
What a regime is
The simplest and most widespread way to label one is where price sits relative to a long moving average. Above it, the market is broadly rising; below it, falling. EMA200 is the usual choice — not because two hundred periods are special, but because it has long been the customary marker and everyone reads it the same way.
It is a crude label, and that is its virtue. It answers "above or below", not "where next", and there is very little to argue about.
What separating by regime gives you
It removes a whole class of false analogues: shape matches that happened under entirely different conditions. Shape is handled by other criteria — this one only answers which market it happened in.
The effect is most visible on long histories. The longer the series, the more stretches it contains that lived through both multi-year rallies and crashes, and the more often a shape from one regime surfaces as an "analogue" for the other.
The short-history problem
A long moving average has an awkward property: at the start of a series it does not exist — there is nothing to compute it from. On an instrument with a short history that cuts off a noticeable part of the data.
There are two ways out, both with costs. Either discard the start of the series entirely and lose data, or use a shorter average there and accept a cruder regime label exactly where data is already scarce. There is no universally correct answer.
A common mistake
Taking the regime label for a signal. "Price is above the EMA200, so the trend is up, so buy" follows from nothing: a moving average is computed from past prices and lags by definition.
In analogue search it has a different job — not to predict, but to separate the incomparable.
What separating does not give you
It does not make the surviving matches good. It only removes the ones that were plainly about something else. Every other caveat about analogue statistics — small samples, survivorship in the data, dependence on the threshold — still applies.